Hybrid attention-based deep learning for multi-label ophthalmic disease detection on fundus images

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Ophthalmic diseases severely affect vision and quality of life.
• Early diagnosis via fundus images is crucial for effective treatment.
• Existing deep learning models struggle with accuracy, interpretability, and efficiency in multi-label ophthalmic disease classification.

Purpose of the Study:

• To introduce HAM-DNet, a novel hybrid deep learning model for multi-label ophthalmic disease detection.
• To enhance diagnostic accuracy and interpretability in automated fundus image analysis.
• To develop an efficient model suitable for clinical deployment.

Main Methods:

• HAM-DNet integrates EfficientNetV2 and Vision Transformers (ViT).
• Attention-based feature refinement is achieved using Squeeze-and-Excitation (SE) blocks.
• A U-Net-based lesion localization module improves interpretability.
• The model was trained and validated on diverse fundus image datasets (ODIR-5K, Messidor, G1020, JSIEC).

Main Results:

• HAM-DNet demonstrated high performance: 95.3% accuracy, 96.2% precision, 97.1% recall, 98.42 AUC, and 96.75 F1-score.
• The model achieved a low computational cost of 9.7 GFLOPS.
• HAM-DNet surpassed existing models like Shallow CNN and EfficientNet in multi-label classification accuracy and reducing diagnostic errors.

Conclusions:

• HAM-DNet provides a reliable, accurate, and interpretable method for detecting multiple ophthalmic conditions.
• Its lightweight design facilitates clinical integration, particularly in telemedicine and low-resource settings.
• The model supports automated, efficient, and precise diagnosis of eye diseases.